Version 1.1 of ai05s/ai05-0050-1.txt

From: Stephen W. Baird
Sent: Friday, April 20, 2007  2:44 PM

!question

Should the implementation permissions of 6.5(24/2) be extended?

!discussion

The implementation permissions of 6.5(24/2) do not go far enough.

This section states:
  If the result subtype of a function is unconstrained, and a call on the
  function is used to provide the initial value of an object with a
  constrained nominal subtype, Constraint_Error may be raised at the point
  of the call (after abandoning the execution of the function body) if,
  while elaborating the return_subtype_indication or evaluating the
  expression of a return statement that applies to the function body, it
  is determined that the value of the result will violate the constraint
  of the subtype of this object.
  
This is a fine start, but more is needed in order to efficiently
implement initialization in place as required by AI95-00318-02 .

--------

Issue #1:

If an object is being initialized by a function call and the implementation
can tell before the call that the object subtype and the function result
subtype are incompatible, then the implementation should be permitted
to raise an exception before the call.

Consider:

  declare
    type T;
    type Inner (D : access T) is limited null record;
    type T is
	limited record
	    F : Inner (Outer'Access);
	end record;
       -- requires initialization in place

    type Unconstrained is array (Positive range <>) of T;
    
    subtype S1 is Unconstrained (1 .. Ident_Int (10));
    subtype S2 is Unconstrained (1 .. Ident_Int (11));

    function F return S1 is
    begin
        Text_Io.Put_Line ("F was called");
        return (others => <>);
    end F;

    X : S2 := F;
  begin null; end;

An implementation should be allowed to raise Constraint_Error before the
call to F (without invoking 11.6).

In the case where the result subtype of a function is a constrained definite
structure, an implemention ought to be able to have a caller
simply pass in a buffer address; nothing more complicated than that should
be needed.

In order to implement initialization in place, F must be passed the address of
X. An implementation must verify before the call that S1 and S2 have compatible
constraints - it would not do to pass in, for example, the address of a small
object to a function which is going to fill in a large result.

The question is what to do if that test fails. The simple thing would be to
raise Constraint_Error at that point. The annoyingly complicated thing that
the language currently seems to require is to allocate a buffer of the size
the function expects, pass that buffer address into the function, finalize the
returned value, and then raise Constraint_Error (or something like that).

One could go further and handle cases involving qualified expressions
such as

    Y : S1 := S2'(F);

, but this doesn't seem necessary.

--------

Issue #2:

Initialization in place is only required for limited types, so it would
be ok for an implementation to ignore the implementation permissions of
6.5(24/2) for a nonlimited type. If an implementation chooses not to
do this, however, there can be problems with extended return statements
which attempt to modify discriminants of the function result object after
it has been initialized.

Consider:

  subtype Index is Integer range 1 .. 20;

  type Rec (Length : Index := 10) is
    record
      F : String (1 .. Length);
    end record;

  function F return Rec is
  begin
    return Result : Rec do
      Result := (5, "12345");
      Text_Io.Put_Line ("Assignment succeeded");
    end return;
  end F;

  X1 : Rec (10) := F;

For an in-out or out mode parameter with an unconstrained nominal subtype, 
the formal inherits the constraints (if any) of the actual. The same
model should apply, at the discretion of the implementation, to return objects
in the case where the nominal subtype of the return object is unconstrained
and definite and "a call on the function is  used to provide the initial value
of an object with a constrained nominal subtype".

This means that for the preceding example, execution of F could (at the
discretion of the implementation) be abandoned (and Constraint_Error raised at
the call site) at the point of the assignment statement. If the literal
"10" in the declaration of X1 were replaced with some other value, then the
default initialization of Result could raise Constraint_Error.

There is also a similar issue in the case of a function call which is
used to initialize an allocator when the designated subtype of the
access type is unconstrained but allocated objects themselves are constrained.

-------

Issue #3:

The implementation permissions of 6.5(24/2) should apply recursively in
the case where a function call returns the result of another function call.

Consider:

  declare
    type T;
    type Inner (D : access T) is limited null record;
    type T is
	limited record
	    F : Inner (Outer'Access);
	end record;
       -- requires initialization in place

    type Unconstrained is array (Positive range <>) of T;
    
    function Some_Function return T is ... ;

    function F1 return Unconstrained is
    begin
      return (1 .. 10 => Some_Function);
    end F1;

    function F2 return Unconstrained is
    begin
      return F1;
    exception
      when Constraint_Error =>
        return (1 .. 3 => <>);
    end F2;

    X : Unconstrained (1..3) := F2;
  begin null; end;
  

In the execution of this example, it should be ok for the return statement
in F1 to raise Constraint_Error before executing any calls to Some_Function.

One might argue that this freedom is already implicit in the current wording,
but this does not seem to be obvious.

--------

It seems that it would not make sense to propose RM wording for these issues
until we reach consensus on which, if any, of these issues require
any action at all.

One could argue that no action is required for any them:
   #1 is essentially a complaint that the language requires an
      implementation to do something that is silly and complex, but
      not something that is undefined or impossible.
   #2 is only a concern for implementations which choose to implement
      initialization in place for functions with unconstrained definite
      non-limited discriminated result subtypes. The language only
      requires initialization in place in the limited case.
      If an implementation doesn't want to deal with the complications of
      implementing the given example, then maybe it shouldn't choose to
      implement initialization in place for this case.
   #3 could be viewed as requiring only a clarification of the current
      wording.

****************************************************************

From: Tucker Taft
Sent: Friday, April 20, 2007  4:05 PM

...
> In the case where the result subtype of a function is a constrained definite
> structure, an implemention ought to be able to have a caller
> simply pass in a buffer address; nothing more complicated than that should
> be needed.
> 
> In order to implement initialization in place, F must be passed the address of
> X. An implementation must verify before the call that S1 and S2 have compatible
> constraints - it would not do to pass in, for example, the address of a small
> object to a function which is going to fill in a large result.

I had presumed that for cases like this, one would use
an implementation based on the existing "OUT" parameter model,
where there is a "constrained" bit and an object whose
discriminants are of interest if the object is
constrained.  The function would check the constrained bit,
and if True, check to be sure the discriminants match.
If they don't match, it would propagate an exception.

On the other hand, for OUT parameters, there is a check
before the function is called if the formal subtype is
constrained, so one could argue that that is a precedent
for raising the exception before the call occurs if
the return subtype is constrained and so is the target
object, and the subtypes don't match.

> The question is what to do if that test fails. The simple thing would be to
> raise Constraint_Error at that point. The annoyingly complicated thing that
> the language currently seems to require is to allocate a buffer of the size
> the function expects, pass that buffer address into the function, finalize the
> returned value, and then raise Constraint_Error (or something like that).

Where does the language require this?  If the model is
"initialize in place," then allocating a buffer the size
the function "expects" is clearly inappropriate.  I think
the language presumes you never allocate "extra" space
when calling these functions.  You might be assigning
the function call into a component of another object,
so the space might already "exist" in that sense.
So the function can't "expect" a space of a particular
size, it must deal with what is the nominal subtype
of the target object.

In any case, I like your suggested generalization,
but in the absence of that, functions returning constrained
subtypes of limited types whose first subtype is unconstrained
must deal with the possibility that the target object has
the wrong constraints.  Note that this applies to functions
returning arrays as well.  It certainly seems annoying to
have to pass in a 'constrained bit when the result subtype
is constrained.  There is no precedent for having to do
something like that for arrays.

> One could go further and handle cases involving qualified expressions
> such as
> 
>     Y : S1 := S2'(F);
> 
> , but this doesn't seem necessary.

I would think if we allow the optimization, then we
should be fairly flexible about it.

,,,
> There is also a similar issue in the case of a function call which is
> used to initialize an allocator when the designated subtype of the
> access type is unconstrained but allocated objects themselves are constrained.

I think a nastier situation occurs when the default
value of the discriminant does *not* match that of the
target object, but the assignment produces a value
that *does* match.  For example:

    X1 : Rec(5) := F;

This is a case where if you do the optimization,
constraint_error is raised.  If you don't do the
optimization, then the function call succeeds
and X1 is appropriately initialized.  I think this
means that for nonlimited types, if the result
object is unconstrained, you must *not* do the
optimization on the off chance that the final
value of the discriminants at the point of the
actual return to the caller *do* match the target's
nominal subtype.

> -------
> 
> Issue #3:
> 
> The implementation permissions of 6.5(24/2) should apply recursively in
> the case where a function call returns the result of another function call.

I always presumed that they did.  Clarification should
be provided if the words imply otherwise.

> ...
> It seems that it would not make sense to propose RM wording for these issues
> until we reach consensus on which, if any, of these issues require
> any action at all.
> 
> One could argue that no action is required for any them:
>    #1 is essentially a complaint that the language requires an
>       implementation to do something that is silly and complex, but
>       not something that is undefined or impossible.

It is pretty bad for functions returning arrays.
I think your generalization should be allowed.

>    #2 is only a concern for implementations which choose to implement
>       initialization in place for functions with unconstrained definite
>       non-limited discriminated result subtypes. The language only
>       requires initialization in place in the limited case.
>       If an implementation doesn't want to deal with the complications of
>       implementing the given example, then maybe it shouldn't choose to
>       implement initialization in place for this case.

I think we need to *disallow* the optimization for nonlimited
types when the return object is unconstrained.  It seems
very bad if the optimization results in a Constraint_Error
that wouldn't occur if you implemented the "canonical"
semantics.

>    #3 could be viewed as requiring only a clarification of the current
>       wording.

I hope it is just a clarification, but if we are going to have
to reword this section a bit anyway, we might as well be
explicit about nested calls.

****************************************************************

From: Randy Brukardt
Sent: Friday, April 20, 2007  5:08 PM

> I think a nastier situation occurs when the default
> value of the discriminant does *not* match that of the
> target object, but the assignment produces a value
> that *does* match.  For example:
>
>     X1 : Rec(5) := F;
>
> This is a case where if you do the optimization,
> constraint_error is raised.  If you don't do the
> optimization, then the function call succeeds
> and X1 is appropriately initialized.  I think this
> means that for nonlimited types, if the result
> object is unconstrained, you must *not* do the
> optimization on the off chance that the final
> value of the discriminants at the point of the
> actual return to the caller *do* match the target's
> nominal subtype.

I don't get it (although I didn't read this very carefully). If you have a
constrained object, you have to pass in those constraints with the object.
(That is, initialization of constraints and of the rest of the object might
be done separately, and in separate places.) Indeed, this example is pretty
much the *only* case where I'd even consider doing the optimization. (I'd
never consider it for unconstrained types, nor for assignments; it's just
too messy in those cases.) If it can't be done here, then it is worthless
for non-limited types (at least for those with any sort of discriminants or
bounds), and then why are we talking about it at all?

****************************************************************

From: Tucker Taft
Sent: Friday, April 20, 2007  5:35 PM

I would disallow the optimization if the object declared
in the extended return statement is *unconstrained* and
nonlimited.  This means that its discriminant values might
change before the extended return statement finishes.
I would allow it if the return object is constrained by
its initial value, or if it is nominally constrained.

****************************************************************

From: Stephen W. Baird
Sent: Friday, April 20, 2007  6:05 PM

> > The question is what to do if that test fails. The simple thing would
> > be to raise Constraint_Error at that point. The annoyingly complicated
> > thing that the language currently seems to require is to allocate a
> > buffer of the size the function expects, pass that buffer address into
> > the function, finalize the returned value, and then raise Constraint_Error
> > (or something like that).
> 
> Where does the language require this? 

You're right, it really doesn't.

I expressed my opinion that
   "In the case where the result subtype of a function is a constrained
    definite structure, an implemention ought to be able to have a caller
    simply pass in a buffer address; nothing more complicated than that
    should be needed".

I think it follows from this premise that the language effectively
imposes this requirement because there really isn't any other reasonable
alternative available (since we are not currently allowed to raise
the exception before the call).

If you don't accept the premise, then I agree that this behavior is not
required.

> In any case, I like your suggested generalization, ...

... which would allow the simple implementation I described.
We agree on the important point.

> > Issue #2
> 
> I think a nastier situation occurs when the default
> value of the discriminant does *not* match that of the
> target object, but the assignment produces a value
> that *does* match.  For example:
> 
>     X1 : Rec(5) := F;
> 
> This is a case where if you do the optimization,
> constraint_error is raised.  If you don't do the
> optimization, then the function call succeeds
> and X1 is appropriately initialized.  I think this
> means that for nonlimited types, if the result
> object is unconstrained, you must *not* do the
> optimization on the off chance that the final
> value of the discriminants at the point of the
> actual return to the caller *do* match the target's
> nominal subtype.

Good point.

I agree with your recommendations regarding all three issues.

****************************************************************